Since the time of the MCEV models in the early 2000’s European life insurance companies have been facing a computational challenge: How to accurately project the entire portfolio consisting of thousands of policies with the existing IT capacities? The models become ever more complex, due to regulatory and internal company requirements. Moreover, the entire portfolio must be projected far into the future, in case of life-long contracts often 60 years or more. To make matter worse, there is no analytical expression for the best estimate of the portfolio value (own funds from shareholders’ perspective and best estimate liabilities from policyholders’ perspective). Therefore, companies have to perform thousands of probability weighted Monte Carlo simulations. All this taken together leads to exploding runtimes even on large, expensive in-house server farms or with an efficient usage of cloud technology.
Although the so-called Moore’s Law, according to which the computational capacity doubles every two years, has maintained its validity during the last two decades, this was not enough for the life insurers to be able to meet the deadlines and keep the IT costs under control. In most cases, the runtime of cash flow projection models greatly exceeds the available time for Solvency II or IFRS closure, if companies attempt to project each policy separately.
The solution the companies have developed over the years consists in identifying and grouping similar policies and projecting only one representative of each group. This clustering of policies is known as grouping or optimization of the portfolio.
In the webinar we will show various methods for grouping of life insurance policies, starting from the more traditional approaches of non-negative least-squares optimization, over the methods of unsupervised machine learning like k-means clustering and supervised learning like neural networks, all the way to the most recent methods leveraging on the simplex methods from mathematical programming. This is an active research area and the latest successes in developing efficient methods without a need of manual intervention will also be presented.
How these methods perform on different realistic portfolios of policies will be another important part of the webinar.